What happened
On PR #2246, the code agent implemented conflict-resolution guidance using spec-drift as the category name, matching issue #2245's proposal. The human reviewer (ralphbean) explicitly requested renaming it to spec-change and requiring spec updates in the same PR. The fix agent correctly applied these changes. However, the second review run raised a medium-severity spec-change finding stating the implementation deviated from issue #2245's proposed naming (spec-drift vs spec-change), recommending the PR either align with the issue or document the rationale. This is a false positive — the human had explicitly authorized the rename in their review.
What could go better
The intent-coherence sub-agent evaluates scope authorization solely against the linked issue. It has no mechanism to incorporate human review comments from earlier in the PR lifecycle. When a human requests changes that deviate from the issue's original specification, the review agent cannot distinguish 'human-authorized scope change' from 'unauthorized scope creep.' This is a confirmed gap — the fix-review skill explicitly states 'human instruction supersedes bot review feedback,' but the review agent on the next pass has no way to access or respect that human instruction. Confidence: high. The false positive is clearly documented in the review output, and the architectural gap is confirmed by reading both the intent-coherence sub-agent definition and the review orchestrator's SKILL.md (neither has any mechanism for reading prior human review comments). Related issues #664 and #1273 address broader aspects of incorporating context, but neither specifically addresses the intent-coherence sub-agent's scope-authorization logic treating human-directed changes as deviations.
Proposed change
Modify the review orchestrator in internal/scaffold/fullsend-repo/skills/pr-review/SKILL.md to extract human review comments (not just prior bot findings) during the prior-review anchoring step (step 2a) and pass them to sub-agents as additional context. Specifically for the intent-coherence sub-agent (sub-agents/intent-coherence.md), add guidance in the scope-authorization section: when a human reviewer has explicitly requested a change that deviates from the linked issue's specification, the intent-coherence sub-agent should treat the human's review as an amendment to the authorized scope. The sub-agent should still note the deviation (as an info-level finding so the issue can be updated), but should not raise it as a medium+ finding that implies unauthorized scope creep. Cross-reference #664 for the broader human-review incorporation work.
Validation criteria
On the next 5 PRs where a human requests changes via review that deviate from the linked issue's original specification, the review agent's post-fix review should not raise medium+ findings about the deviation from the issue. The deviations should appear as info-level findings at most. Verify by checking whether any medium+ intent-coherence findings reference 'issue specification' or 'authorized scope' for changes that a human reviewer explicitly requested.
Generated by retro agent from #2246
What happened
On PR #2246, the code agent implemented conflict-resolution guidance using
spec-driftas the category name, matching issue #2245's proposal. The human reviewer (ralphbean) explicitly requested renaming it tospec-changeand requiring spec updates in the same PR. The fix agent correctly applied these changes. However, the second review run raised a medium-severityspec-changefinding stating the implementation deviated from issue #2245's proposed naming (spec-driftvsspec-change), recommending the PR either align with the issue or document the rationale. This is a false positive — the human had explicitly authorized the rename in their review.What could go better
The intent-coherence sub-agent evaluates scope authorization solely against the linked issue. It has no mechanism to incorporate human review comments from earlier in the PR lifecycle. When a human requests changes that deviate from the issue's original specification, the review agent cannot distinguish 'human-authorized scope change' from 'unauthorized scope creep.' This is a confirmed gap — the fix-review skill explicitly states 'human instruction supersedes bot review feedback,' but the review agent on the next pass has no way to access or respect that human instruction. Confidence: high. The false positive is clearly documented in the review output, and the architectural gap is confirmed by reading both the intent-coherence sub-agent definition and the review orchestrator's SKILL.md (neither has any mechanism for reading prior human review comments). Related issues #664 and #1273 address broader aspects of incorporating context, but neither specifically addresses the intent-coherence sub-agent's scope-authorization logic treating human-directed changes as deviations.
Proposed change
Modify the review orchestrator in
internal/scaffold/fullsend-repo/skills/pr-review/SKILL.mdto extract human review comments (not just prior bot findings) during the prior-review anchoring step (step 2a) and pass them to sub-agents as additional context. Specifically for the intent-coherence sub-agent (sub-agents/intent-coherence.md), add guidance in the scope-authorization section: when a human reviewer has explicitly requested a change that deviates from the linked issue's specification, the intent-coherence sub-agent should treat the human's review as an amendment to the authorized scope. The sub-agent should still note the deviation (as an info-level finding so the issue can be updated), but should not raise it as a medium+ finding that implies unauthorized scope creep. Cross-reference #664 for the broader human-review incorporation work.Validation criteria
On the next 5 PRs where a human requests changes via review that deviate from the linked issue's original specification, the review agent's post-fix review should not raise medium+ findings about the deviation from the issue. The deviations should appear as info-level findings at most. Verify by checking whether any medium+ intent-coherence findings reference 'issue specification' or 'authorized scope' for changes that a human reviewer explicitly requested.
Generated by retro agent from #2246